Artificial Intelligence

From fundamentals to frontier models

Artificial intelligence is the study and engineering of systems that perform tasks associated with intelligence: perceiving, predicting, reasoning, generating language and images, and acting toward goals. Almost all practical AI today is machine learning, where behavior is learned from data instead of written as rules. Most of that is deep learning, and the systems in the headlines are large foundation models trained on broad data and then adapted to many tasks.

This hub is the entry point to the site’s AI material. It covers the theory at four levels of depth, from a no-math introduction to graduate-level learning theory, the individual subfields, and the hands-on generative-AI guides. It also gives a short account of how modern models are built and where the field stands in late 2026.

How the field nests

The terms AI, machine learning, deep learning, and generative AI are often used interchangeably, but they name nested sets. Each inner set is a specific approach within the one around it:

flowchart TB
    subgraph AI["Artificial intelligence: search, logic, planning, learning"]
        subgraph ML["Machine learning: behaviour learned from data"]
            subgraph DL["Deep learning: multi-layer neural networks"]
                FM["Foundation models<br/>LLMs, vision-language models,<br/>image and video generators"]
            end
        end
    end

Symbolic AI (search, logic, knowledge representation, planning) is the part of AI outside machine learning. It still underpins route planning, game-tree search, constraint solvers, and formal verification, and it increasingly shows up inside learned systems as tools an LLM calls.

Reading levels

The core theory is written at four depths. Each builds on the one before it, but you can stop at whichever level fits your goal.

Level Page Audience Covers
1 AI Fundamentals: Simplified Anyone; no math What learning from data means, neural networks by analogy, what LLMs and image generators do
2 AI & Machine Learning Engineers; some calculus and linear algebra The full technical reference, split into focused sub-pages (listed below)
3 AI Deep Dive Practitioners Transformers, large language models, and research directions
4 Advanced AI Mathematics Graduate level Statistical learning theory, optimization, generalization bounds, and proofs

The hands-on material lives in the AI/ML guides, which focus on image generation with diffusion models (Stable Diffusion, SDXL, SD3, FLUX, ComfyUI, LoRA training, ControlNet) along with MLOps and model compression.

flowchart LR
    Hub["AI hub<br/>(this page)"] --> L1["1. Simplified"]
    L1 --> L2["2. AI & ML reference"]
    L2 --> L3["3. Deep Dive"]
    L2 --> L4["4. Mathematics"]
    L2 --> Sub["Sub-pages: foundations,<br/>architectures, generative,<br/>RL, fine-tuning, ethics"]
    Hub --> Guides["AI/ML guides"]
    Guides --> Img["Diffusion: SD, SDXL,<br/>SD3, FLUX, ComfyUI"]
    Guides --> Train["LoRA, ControlNet,<br/>fine-tunes"]
    Guides --> Prod["MLOps, compression,<br/>pipelines"]
    L3 --> QML["Quantum ML<br/>(Quantum Computing hub)"]

Subfields

AI is several overlapping fields, distinguished mostly by the kind of data they work with and the structure they exploit. Since the late 2010s the transformer architecture has pulled most of them together: the same basic model family now handles text, images, audio, video, and actions.

Subfield Problem Characteristic methods On this site
Machine learning Learn a function from examples to predict or decide Linear models, decision trees and gradient boosting, SVMs, clustering, dimensionality reduction ML Foundations, Core ML Algorithms, Loss Functions
Deep learning Learn hierarchical representations from raw input Backpropagation, CNNs, RNNs, transformers, normalization, residual connections Deep Learning Architectures, Deep Learning Theory
Natural language processing Understand and generate human language Tokenization, pretrained transformers, LLMs, retrieval-augmented generation AI Deep Dive
Computer vision Interpret images and video CNNs, vision transformers, detection and segmentation models, vision-language models ControlNet, Inpainting & Editing
Generative modeling Sample new data from a learned distribution Autoregressive models, diffusion and flow matching, VAEs, GANs Generative Models, Stable Diffusion, FLUX
Reinforcement learning Learn to act by trial, error, and reward MDPs, Q-learning, policy gradients, PPO, RL from human or verifiable feedback Reinforcement Learning, Game AI
Adaptation & deployment Specialize and serve pretrained models efficiently Fine-tuning, LoRA/PEFT, quantization, distillation, MLOps Fine-Tuning, LoRA Training, Model Compression, MLOps

The main learning paradigms cut across all of these subfields:

Paradigm Training signal Examples
Supervised Labeled input/output pairs Image classification, spam filtering, regression
Unsupervised Structure in unlabeled data Clustering, PCA, density estimation
Self-supervised Labels derived from the data itself (predict the next token, the masked patch, the added noise) LLM and diffusion pretraining
Reinforcement Scalar reward from an environment or evaluator Game playing, robotics, RLHF, reasoning-model training

How a modern foundation model is built

Most of the models discussed on this site, whether language models or image generators, go through the same broad stages. Understanding them explains much of their behavior, including their strengths, failure modes, and cost profile.

flowchart LR
    Data["Web-scale data<br/>(text, code, images,<br/>audio, video)"] --> Pre["Pretraining<br/>self-supervised,<br/>next-token or denoising"]
    Pre --> Base["Base model"]
    Base --> Post["Post-training<br/>SFT, preference tuning<br/>(RLHF, DPO), RL with<br/>verifiable rewards"]
    Post --> Model["Assistant / reasoning<br/>model"]
    Model --> Inf["Inference<br/>sampling, test-time<br/>reasoning, quantized serving"]
    Inf --> Sys["System<br/>RAG, tools, agents,<br/>guardrails, evals"]
  1. Pretraining. A large network, usually a transformer and often a sparse mixture-of-experts, is trained with a self-supervised objective on a very large corpus: predict the next token for language models, or predict the noise or velocity for diffusion and flow-matching image models. Scaling laws (Kaplan et al., 2020; the “Chinchilla” results of Hoffmann et al., 2022) showed that loss falls predictably as parameters, data, and compute grow together. Much of the compute and almost all of the model’s knowledge comes from this stage.
  2. Post-training. Supervised fine-tuning on curated demonstrations teaches the model to follow instructions. Preference optimization (reinforcement learning from human feedback, or direct preference optimization) steers it toward responses people prefer. Since 2024, large-scale reinforcement learning on verifiable tasks such as mathematics with checkable answers and code with tests has produced reasoning models that generate long chains of intermediate steps before answering.
  3. Inference. Output quality now also depends on test-time compute: letting a model think longer, sample several candidates, or check its own work can improve accuracy on hard problems, trading latency and cost for quality. Serving costs are reduced with quantization, distillation into smaller models, speculative decoding, and caching.
  4. Systems around the model. Production applications seldom call a bare model. They add retrieval-augmented generation (RAG) to ground answers in current or private data, tool use to act on external systems, agent loops that plan and execute multi-step tasks, and evaluation and guardrail layers to measure and constrain behavior.

The same stages apply to image models. A diffusion model is pretrained to denoise, fine-tuned for aesthetics or instruction following, adapted with LoRA or ControlNet, and served with fewer sampling steps through distillation. The AI/ML guides cover these steps in practice.

State of the field (late 2026)

Milestones

timeline
    title Selected milestones in modern AI
    2012 : AlexNet wins ImageNet, starting the deep-learning era
    2016 : AlphaGo defeats Lee Sedol
    2017 : Attention Is All You Need introduces the transformer
    2020 : GPT-3 shows in-context learning at scale : DDPM revives diffusion models : AlphaFold 2 at CASP14
    2022 : Stable Diffusion released with open weights : ChatGPT launches (November)
    2023 : GPT-4 and multimodal LLMs : open-weight LLMs become competitive
    2024 : Nobel prizes for Hopfield and Hinton (Physics) and for Hassabis, Jumper and Baker (Chemistry) : OpenAI o1 introduces RL-trained reasoning models : EU AI Act enters into force : Model Context Protocol released
    2025 : DeepSeek-R1 open-weight reasoning model : EU AI Act obligations for general-purpose models apply : MCP moves to the Linux Foundation's Agentic AI Foundation
    2026 : EU Digital Omnibus defers high-risk AI Act obligations to 2027-2028

Current directions

  • Reasoning models and test-time compute. Following OpenAI’s o1 (September 2024) and the open-weight DeepSeek-R1 (January 2025), which published its reinforcement-learning recipe, every major lab now ships models that reason at length before answering. Many let the caller set a thinking budget. Progress on mathematics, science, and coding benchmarks has been rapid enough that many older benchmarks are saturated, and evaluation has shifted toward harder expert-written sets and long, realistic tasks.
  • Agents and tool use. Models increasingly operate software instead of only producing text: calling APIs, writing and running code, browsing, and working over multi-hour tasks. The Model Context Protocol (released by Anthropic in November 2024 and donated to the Linux Foundation’s Agentic AI Foundation in December 2025) has become the common interface between models and tools. The main open problems are reliability over long horizons, security (especially prompt injection through untrusted content), and evaluation.
  • Multimodality. Frontier models take text, images, audio, and video as input. Generation of images, speech, and increasingly video with synchronized audio has become routine. In image generation, diffusion transformers trained with flow matching (such as SD3 and FLUX) have largely replaced U-Net latent diffusion.
  • Open-weight models. Open-weight language and image models trail the closed frontier by a relatively short margin, which makes local deployment, fine-tuning, and research on model internals practical.
  • Efficiency. Mixture-of-experts architectures, low-bit quantization, distillation, and better inference kernels have steadily lowered the cost of a given capability level, and capable models now run on laptops and phones.
  • Safety, alignment, and interpretability. Work covers training models to follow specified behavior, evaluating dangerous capabilities before release, and mechanistic interpretability, which tries to identify the internal features and circuits a network uses. See Frontier Research & Ethics.
  • Regulation. The EU AI Act entered into force in August 2024, with obligations for general-purpose AI models applying from August 2025. The Digital Omnibus on AI (in force July 2026) deferred the high-risk system obligations to December 2027 for stand-alone systems and August 2028 for AI embedded in regulated products, while the Article 50 transparency obligations still applied from August 2026.
  • AI for science. Following AlphaFold’s impact on protein-structure prediction, learned models are widely used in materials discovery, weather forecasting, mathematics, and as research assistants.

Specific model names and benchmark leaders change every few months, so this hub deliberately leaves them out. The sub-pages describe architectures and techniques, which age more slowly.

Learning paths

Goal Path
Understand how AI works Simplified → AI & ML reference → ML Foundations → Deep Learning Architectures → AI Deep Dive
Generate images and train models Stable Diffusion Fundamentals → ComfyUI → Model Types → Base Models Comparison → LoRA Training → Advanced Techniques
Ship ML to production Fine-Tuning → Model Compression → Optimization & Performance → MLOps & Production → Production Pipelines
Research and theory AI & ML reference → Deep Learning Theory → Advanced AI Mathematics → Quantum Computing (quantum ML)
Area Pages
Infrastructure Docker, Kubernetes, AWS, CI/CD
Serving models behind APIs API Design, Distributed Systems, Observability
Theoretical connections Statistical Mechanics (energy-based models, the physics behind diffusion), Quantum Computing

See also